What AI model does it use?

What AI model does Siri use?

Apple shipped its rebuilt assistant, Siri AI, on September 14, 2026, in beta and in English first. The question of what runs it has a two-part answer. The models are Apple's, with Apple's names on them. But Apple built them with Google, and it says so.

The Google deal

The arrangement was announced in a joint statement on January 12, 2026: "the next generation of Apple Foundation Models will be based on Google's Gemini models and cloud technology." The same statement said Apple Intelligence would "continue to run on Apple devices and Private Cloud Compute", Apple's own server system for requests too big for a phone.

Apple's September launch release describes the result the same way. Siri AI's new abilities "are powered by the next generation of Apple Foundation Models, custom-built in collaboration with Google and its Gemini models", and those models "run on device and on servers using Private Cloud Compute."

The five Apple models behind it

Apple's machine learning team published the details on June 8, 2026. The third generation, AFM 3, is "a family of five foundation models custom-built in collaboration with Google":

That last row surprised me. Four of the five are "purpose-built for Apple silicon", but for Cloud Pro, Apple writes, "we worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud", with the same privacy guarantees. So the biggest Siri model runs physically inside Google's data centers, on Apple's software.

Where ChatGPT fits now

ChatGPT hasn't gone away, but it's no longer anything like the engine. Apple's support guide treats it as an opt-in extension: "If you're using Siri AI (Beta), tap Set Up Extensions, tap ChatGPT, then turn on Use Extension." An OpenAI account is optional, needed only if you want your ChatGPT history kept.

Developers get their own choice, which is easy to confuse with Siri's. Apple's Foundation Models framework lets apps use Apple's models or, in Apple's words, "models of their choice, like Claude and Gemini". That choice belongs to app makers using Apple's tools; Siri itself doesn't offer it.

Free, with limits

The features that need Apple's servers come with daily usage limits, and Apple's Apple Intelligence page says expanded access "will be available for a fee in the future." It works on the iPhone 15 Pro and Pro Max and on iPhone 16 models or later. Siri AI isn't initially available in the EU on iPhone, iPad and Apple Watch, or in China.

What Apple doesn't spell out

Apple hasn't said which Gemini models AFM 3 came from, or what "in collaboration with Google" meant in practice (whether Apple trained on Gemini's outputs, fine-tuned a Gemini model, or something else). It also doesn't say which of the five models answers a particular Siri request, or how big the server models are. Numbers you may see elsewhere, like a trillion-parameter Gemini or a yearly fee, come from press reports rather than Apple or Google.